The internal C++ constructors accept per-instance payload capacities. Ordinary callers keep the existing payload defaults: 64 KiB for position-tail pages, 64 successor-root rows, 64 group-effect receipt rows, 16 long-chain index rows, and 2,730 complete 24-byte curve points. These read caches use detail::ReadThroughCache, with direct mapping and a read-only backing port. Owner writes retain their existing ordering and invalidate read tags before changing backing. The retained-sequence backing keeps its separate 64 KiB payload default for its shared resident-prefix/read-cache allocation.

These capacities change resident storage and backing traffic. They do not change retained history, identifier widths, command order, logical record counts, curve points, or report mode. strategy_set_cache_policy_v1 and BacktestEngine::set_cache_policy supply inclusive byte allowances to the admitted run. The payload/row constructor overloads remain available for standalone C++ owners.

PositionTailSpill streams both its 30-byte backwards footers and arbitrary-length entry identifiers across pages of at most 4,096 bytes. Allowances below one page select that smaller geometry; larger allowances keep complete 4,096-byte pages. Its fixed transfer buffer is 4,096 bytes. The final page reads only its existing prefix. Zero and optional-allocation failure read backing through the shared cache's bypass path. Append and clear invalidate tags before I/O. The footer codec and newest-first traversal order are unchanged. The typed CacheCapacity overload uses fixed 4,096-byte pages; its default is 16 complete shared lines, preserving the 64 KiB payload. The host never converts that allowance using 64 KiB geometry.

WorkingRequestCore accepts TailWindowConfig at construction. The existing one-argument constructor keeps the original policy. Move and reset preserve the selected policy; the lazy stores receive it when they are created. Its native stores retain their authoritative pending batches: replacing these with evictable read lines would put writes into nonthrowing installation or change failed-write retry ordering. Preparation still flushes half of the receipt/chain batch or the full root batch. Zero and pending-allocation failure retain one mandatory pending record per store. Preparation spills that record before another nonthrowing installation. Committed rows use shared read caches with the selected row capacities. Pending storage and read-cache storage coexist and are reported separately. The row setting is not an inclusive byte allowance for their sum. The typed WorkingRequestCore overload instead accepts three independent CacheCapacity values for committed root, receipt and chain-index reads. Host policy uses this overload. The pending batches keep the internal 64/64/16-row defaults at every host cache size, with the same allocation fallback and prepare-only flush rules. Zero disables the read cache without changing pending workspace, retained rows or slab admission. Move and reset preserve both policies in the opaque journal owner.

TailLog accepts absolute indices in [base(), size()). A read below base() is a caller error following explicit history retirement. It keeps the existing out_of_range error. A retained row that has left the resident tail is still read from backing. Cache replacement never advances base(). Backing I/O, corrupt records, slab capacity and arena capacity keep their existing failures.

The retained sequence index is derived from retained placements. Its constructor forwards one typed allowance to the command-backing factory, preserving the 512-byte geometry and two mandatory transaction slots. PlacementTable retains that typed value until lazy index construction. The factory derives complete payload pages once, then uses the same slab for both the initial disk prefix and the shared read cache after spill. Zero bypasses both phases. The arena branch disables the prefix and uses the selected read cache. The existing placement constructor with two byte arguments still accepts payload budgets, adding the exact shared line charge once. The first placement defers index creation; a later mutation or reader may build it. Configuration does not alter rebuilding or query ordering.

ScratchCurveWindow rounds its payload allowance down to complete 24-byte point pages. Backing misses adjacent to the previous lookup fill aligned blocks of up to 2,730 points; other backing misses read one point. Zero, a sub-point allowance and optional-allocation failure bypass. A returned point is a copy, so callers can retain it across later reads. A failed read publishes no successful tag and can be retried after backing repair; unrelated valid lines remain valid. The caller still rebinds readers after replacing or clearing a region. Its typed overload consumes the inclusive allowance directly; the ordinary report-sink metric reader receives the admitted equity_curve_pages value through that overload.

The curve writer's 2,730-point pending batch remains mandatory authoritative write workspace. Its failed-flush retry and suffix-replacement contracts remain independent of the read-cache allowance.

For position, curve and native read caches with page size P and payload allowance C, optional shared-cache storage is floor(C/P) * ReadThroughCache::line_bytes(P). The line charge includes all payload, keys, pins and replacement metadata. Fixed owner control and transfer/commit workspace are additional. The position reader defaults to 16 lines of 4,096 bytes; the curve reader defaults to 2,730 lines of 24 bytes, with one metadata charge per point. cache_control_bytes() reports sizeof(ScratchCurveWindow) + sizeof(State), including the fixed 65,520-byte block workspace. Shared-cache statistics charge the optional line tags and payload separately from that control total. Bypass reads do not use the block.

Native pending storage additionally uses rows * record_bytes when the optional batch exists, plus one inline fallback record per store. At the 64/64/16 defaults this is 3,456 optional pending bytes plus 72 inline bytes; the three shared read caches add another 3,456 payload bytes and their line metadata. These extra read caches preserve the pending-write contract but increase default residency. The host allowances include every read-line charge; the fixed pending batches are separate transaction workspace. They must be added when accounting for the complete store. The retained-sequence factory instead reuses one allocation for its initial prefix and subsequent read cache.

Physical read observations are taken at ExtFile::read, successor-root file/slab reads, the position-tail file facade, and read_scratch_points. Raw-descriptor curve reads count actual pread attempts. Sequence-index tests use the factory's independent physical observer at the inner file/arena operation boundary; their cold-read proof counts completed backing reads. Cache hit/miss counts and logical charged-I/O counts cannot establish that a cold read reached backing. Separate read-only cache statistics expose allocated payload and metadata; focused position/curve tests compare these bytes with global allocation hooks. These observations do not establish whole-engine memory flatness. File/arena provider allocations and explicit output projections remain part of the supervisor's complete allocation census.